Hardware-aware comparative study of lightweight convolutional neural networks for Raspberry Pi-based autonomous driving

Abstract

Deploying deep learning models for autonomous driving on resource-constrained edge devices, such as the Raspberry Pi, presents significant challenges due to strict limitations on inference latency and memory capacity. To address these constraints, this study conducts a comprehensive comparative evaluation of lightweight convolutional neural networks (CNNs) optimized for dual-output regression of steering angle and driving speed. We benchmark a task-specific end-to-end baseline (NVIDIA CNN) against representative classification-oriented architectures—including MobileNet, ShuffleNet, EfficientNet, GhostNet, and SqueezeNet—all reformulated for this regression task. Experiments were conducted on a physical Raspberry Pi-based autonomous RC car platform to assess prediction accuracy, inference speed, and real-world closed-loop driving stability using quantitative metrics such as the normalized jerk ratio. Experimental results demonstrate a clear trade-off: while GhostNetV1 0.5x achieved the highest regression accuracy with a Total R² score of 95.8% and MobileNetV1 recorded a competitive MAE of 1.95, they failed to provide stable control due to severe high-frequency steering jitter. Conversely, the NVIDIA CNN proved to be the most practical solution for general edge deployment, achieving the lowest inference latency of 61.1 ms (16.4 FPS) and a minimal memory footprint of 2.78 MB, ensuring stable autonomous navigation (1.50xjerk ratio). Furthermore, ShuffleNetV2 0.5x emerged as the superior architecture for trajectory precision, recording the lowest weighted MAE of 1.60. These findings underscore that theoretical accuracy does not guarantee real-world drivability on embedded systems, providing practical guidelines for hardware-aware model selection in edge-based autonomous driving.